| Popis: |
Herein, the establishment of data-driven prediction and explanation models for three essential process variables in ironmaking blast furnace processes, namely hot metal temperature, silicon concentration, and cooling capacity is demonstrated. Aside a reliable prediction quality of the models with sufficient prediction horizon, an additional main goal has been to establish interpretable and revealing models. To support (linguistic) interpretability, the main focus has been set on the extraction of rule-based models from a large database collected at a particular blast furnace process located at the partner company's site. Due to expected uncertainty in the data, due to, e.g., measurement noise, fuzzy systems are an adequate architectural choice for achieving models in a robust rule-based form. For a fully automatic training of the predictive fuzzy systems new feature ranking methods have been performed on the one side and a special granular rule extraction procedure on the other side. Testing the obtained models on two separate test datasets from consecutive years shows a stable prediction performance without any error drifts and a higher performance than other related machine learning methods including deep neural networks, and others. Moreover, the final models, turned out to have maximal 4–5 inputs and just a couple of rules and allowed to gain new insights into the processes. Version of record |
